A local subspace based nonlinear target detector

Traditional Orthogonal Subspace Projection (OSP) target detection method can not solve the problem of nonlinear mixing of endmember spectra. Meanwhile, Kernelized Orthogonal Subspace Projection (KOSP) method maps the inseparable data into high dimension space where the target endmembers and background endmembers can be separated. However, the background subspace remains the same for different pixels in KOSP, which would lead to false alarms due to the spectral variation. In order to optimize the background subspace and better suppress the false alarms, this paper proposes a local subspace based nonlinear OSP method (LKOSP) for target detection. Kernelization and neighbor spatial information are used to construct variable optimum background projective subspace. In both simulated data and real image experiments, LKOSP showed superior detection performance over other conventional algorithms.